Structured flexibility in recurrent neural networks via neuromodulation

Authors: Julia Costacurta, Shaunak Bhandarkar, David Zoltowski, Scott Linderman

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In empirical experiments, we find that the structured flexibility in the NM-RNN allows it to both train and generalize with a higher degree of accuracy than low-rank RNNs on a set of canonical tasks. Additionally, via theoretical analyses we show how neuromodulatory gain scaling endows networks with gating mechanisms commonly found in artificial RNNs.
Researcher Affiliation Academia Julia C. Costacurta* Stanford University jcostac@stanford.edu Shaunak Bhandarkar* Stanford University shaunakb@stanford.edu David Zoltowski Stanford University dzoltow@stanford.edu Scott W. Linderman Stanford University scott.linderman@stanford.edu
Pseudocode No The paper presents mathematical equations for the model, but it does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes The necessary code required to reproduce the trained models is included in the Supplementary Material. This code will be made available in a public repository upon publication.
Open Datasets Yes In the Measure-Wait-Go (MWG) task (Fig. 3A) [35], the network receives a 3-channel input... We performed our analysis using the four-task set from Duncker et al. [39], which includes the tasks Delay Pro, Delay Anti, Memory Pro, and Memory Anti illustrated in fig. 4A.
Dataset Splits No The paper describes training and testing on different task intervals but does not provide specific numerical splits for training, validation, and testing sets (e.g., percentages or sample counts).
Hardware Specification No The main paper does not explicitly describe the hardware used for experiments. The NeurIPS checklist mentions that compute resources are described in the Supplementary Material, but these details are not in the main text.
Software Dependencies No The main paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes We trained parameter-matched NM-RNNs (N = 100, M = 5, K = 3, τx = 10, τz = 100), LR-RNNs (N = 106, K = 3, τ = 10), vanilla RNNs (N = 31, τ = 10), and LSTMs (N = 15) to reproduce four intervals, then tested their extrapolation to longer and shorter intervals.